Image Processing Apparatus Selective Padding via Heat Map
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating image data for deep learning models often result in over-processing of non-critical regions, leading to decreased learning speed and inference accuracy due to over-learning, as they perform padding processing on all regions equally.
Innovation Solution
An image processing apparatus that identifies and selectively performs padding processing on regions of interest within the image data using a heat map generated by the Grad-CAM method, focusing on areas where the model is truly interested during inference, thereby enhancing the quality of the image data for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If padding processing is performed on all regions of image data, then the machine learning model can process comprehensive information, but the learning speed decreases and inference accuracy deteriorates due to over-learning
Solution Approach 1:
The image data is segmented into regions of interest and non-regions of interest based on heat map analysis. Padding processing is selectively applied only to regions of interest, dividing the processing task into critical and non-critical segments. This segmentation prevents over-processing of irrelevant areas while maintaining necessary processing on important regions, thereby improving learning speed without sacrificing inference accuracy.
Solution Approach 2:
Different processing quality is applied to different regions of the image data. Regions of interest receive full padding processing while non-regions of interest receive reduced or no processing. This local differentiation optimizes the quality-processing balance, concentrating computational resources where they are most needed and avoiding the degradation caused by uniform over-processing across the entire image.
2Quantity of substance
If padding processing is performed on all regions of image data, then comprehensive training data is obtained, but the processing complexity and computational resources increase
Solution Approach 1:
The method extracts and identifies regions of interest from the image data using heat map analysis before applying padding processing. By taking out only the relevant regions for processing, the system reduces the overall processing complexity while still obtaining sufficient training data quantity. This extraction approach eliminates the need to process entire images uniformly, reducing computational burden.
Solution Approach 2:
Instead of applying padding processing to all regions (excessive action), the method applies processing only to necessary regions of interest (partial action). This partial processing approach maintains adequate training data quantity by focusing on critical areas, while significantly reducing processing complexity compared to comprehensive full-image processing.
3Ease of manufacture
If uniform padding processing is applied to all image regions, then the processing method is simple, but the quality of generated training data decreases due to over-processing of non-critical areas
Solution Approach 1:
The method performs preliminary heat map analysis and region identification before applying padding processing. This preliminary action identifies which regions require processing and which do not, enabling subsequent selective processing that improves training data quality. The preliminary step adds minimal complexity but prevents the quality degradation caused by uniform over-processing.
Solution Approach 2:
The processing approach changes from uniform parameter application across all regions to variable parameter application based on region importance. By changing the processing parameters (applying padding only to regions of interest rather than uniformly), the system improves training data quality while maintaining reasonable processing simplicity through automated heat map-based region identification.
Data Source
AI summary
An image processing apparatus, includes a memory; and a processor coupled to the memory and configured to: generate a trained machine learning model by learning a machine learning model using a first set of image data, output an inference result by inputting a second set of image data to the trained machine learning model, and process a region of interest at a time of inference with respect to image data for which an inference result is correct in the second set of image data.